HFGA-Net: High-frequency guided attention network for deepfake detection and localization

The application of deepfake for video and image manipulation has raised severe trust issues and security concerns in modern society. Prior research has revealed that the high-frequency components of forged images exhibit distinctive forensic artifacts. Motivated by this, we propose a novel deepfake detection and localization approach called High-Frequency Guided Attention Network (HFGA-Net), which enhances the detection of forged faces by leveraging high-frequency information to capture fine-grained forgery cues. Specifically, to better extract and encode the fine-grained forgery cues, we first propose the Frequency-Guided Patch-wise Attention (FGPA). The FGPA captures high-frequency information in shallow features. Furthermore, we design the Dynamic High-Frequency Multi-Stage Attention (DHFMA) to capture high-frequency information from multi-stage features. Additionally, to address the critical scarcity of pixel-level annotations for forged regions in existing datasets, we propose a morphology-based mask generation method. Extensive experiments demonstrate that our proposed method exhibits strong generalization across diverse datasets on both detection and localization tasks.